Risk stratification of major arrhythmia events in Japanese patients with Brugada syndrome using machine learning models.

Background: Brugada syndrome (BrS) has been known to cause fatal arrhythmias, and an effective risk stratification method should be developed. Objective: This study aimed to construct a risk prediction model for BrS using machine learning models. Methods: We enrolled 234 Japanese patients with BrS and analyzed the clinical information including the age, gender, history of syncope, family history of BrS or sudden cardiac death, PR interval in lead Ⅱ, QRS duration in V6, RR interval in V1, r-J interval in V1, T-peak-to-T-end interval, max QTc, fragmented QRS, Type-1 in peripheral leads, spontaneous type 1 pattern, aVR sign, presence of early repolarization (ER), and presence of ER in the peripheral leads. We validated the previous stratification method (BRUGADA-RISK and Predicting Arrhythmic evenT [PAT] scores). Next, we constructed 3 machine learning models (logistic regression, support vector machine [SVM], and random forest). To detect the important clinical features, we used SHapely additive exPlanations and constructed a low-dimensional model. Results: The area under the curve (AUC) was 0.57 for the BRUGADA-RISK score and 0.59 for the PAT score. The SVM revealed the highest AUCs. Moreover, the low-dimension model with the SVM (r-J interval in V1, history of syncope, fragmented QRS, presence of ER, T-peak-to-T-end interval, QRS duration in V6, and age) exhibited a higher AUC than the SVM model using all clinical features (mean AUC, 0.77; 95% confidence interval [0.64-0.89], Welch's T-test Conclusion: The machine learning model could be useful for stratifying major arrhythmic events in BrS.